End-to-End Dialog Neural Coreference Resolution: Balancing Efficiency and Accuracy in Large-Scale Systems
Zhang Dong Songhang deng Mingbang Wang Le Dai Jiyuan Li, Xingzu Liu Ruilin Nong · 2025
Read End-to-End Dialog Neural Coreference Resolution: Balancing Efficiency and Accuracy in Large-Scale Systems by Zhang Dong Songhang deng Mingbang Wang Le Dai Jiyuan Li, Xingzu Liu Ruilin Nong (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-05824.txt.
Introduction
The authors present a neural system that links each pronoun and name in a chat to the folk it points to. Such links are a first need for machines that read talk. At big scale, the work must be both fast and right. Their system reads a chat end to end in one pass, with watch tools that weigh the full text at each call.
Speed tricks trim the cost of each read with no loss the team could find. On set test chats, the system beat past ways on grade while keeping its runs quick. Hard tests back its use in live, large gear. The work sets a mark for the next builds in this line, they say. The catalog name field runs the team's names into one string; the head byline, used here, lists them apart.
Catalog record
| CREATOR | Zhang Dong Songhang deng Mingbang Wang Le Dai Jiyuan Li, Xingzu Liu Ruilin Nong |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-05824
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
End-to-End Dialog Neural Coreference Resolution: Balancing Efficiency and Accuracy in Large-Scale Systems Zhang Dong Songhang deng Mingbang Wang Le Dai Jiyuan Li, Xingzu Liu Ruilin Nong Abstract Large-scale coreference resolution presents a significant challenge in natural language processing, necessitating a balance between efficiency and accuracy. In response to this challenge, we introduce an End-to-End Neural Coreference Resolution system tailored for large-scale applications. Our system efficiently identifies and resolves coreference links in text, ensuring minimal computational overhead without compromising on performance. By utilizing advanced neural network architectures, we incorporate various contextual embeddings and attention mechanisms, which enhance the quality of predictions for coreference pairs. Furthermore, we apply optimization strategies to accelerate processing speeds, making the system suitable for real-world deployment. Extensive evaluations conducted on benchmark datasets demonstrate that our model achieves improved accuracy compared to existing approaches, while effectively maintaining rapid inference times. Rigorous testing confirms the ability of our system to deliver precise coreference resolutions efficiently, thereby establishing a benchmark for future advancements in this field. 1 Introduction Efficient coreference resolution systems should balance model size and performance, as seen with solutions like Maverick, which achieves state-of-the-art coreference resolution using only 500 million parameters, outperforming larger models with up to 13 billion parameters Martinelli et al. (2024) . In multilingual contexts, new models based on the CorefUD dataset demonstrate enhanced coreference resolution through various proposed extensions aimed at diverse linguistic features Pražák and Konopík (2024) . Additionally, approaches leveraging coreference resolution can improve understanding in long contexts, as illustrated by the Long Question Coreference Adaptation framework, which helps manage references and organize information effectively Liu et al. (2024a) . The introduction of domain-specific datasets, such as ThaiCoref, enhances coreference resolution for culturally unique languages and phenomena, contributing to more accurate data representation and processing Trakuekul et al. (2024) . These advancements underscore the potential for developing end-to-end systems that maintain both efficiency and accuracy in resolving coreferences across various contexts. However, the development of efficient and accurate coreference resolution systems faces significant challenges. One of the key innovations includes the addition of a singleton detector to enhance performance, which significantly improves model outcomes on benchmark datasets Zou et al. (2024) . The incorporation of sentence-incremental techniques has shown promise by effectively marking mention boundaries and outperforming many current methods Grenander et al. (2023) . In the context of managing long documents, utilizing a dual cache system to separate global and local entity recognition has proven effective in reducing cache misses and improving coreference scores Guo et al. (2023) . Integrating concepts from centering theory into neural models has also demonstrated improvements over state-of-the-art methods, although the gains in performance may be limited due to existing strong pre-trained representations Chai and Strube (2022) ; Jiang et al. (2022) . Additionally, leveraging both heuristic rules and neural models through a hybrid approach can enhance coreference resolution performance by taking advantage of the strengths of each method Wang and Jin (2022a) . However, balancing efficiency and accuracy remains a key issue that needs to be resolved in large-scale coreference systems. We present an End-to-End Neural Coreference Resolution system that prioritizes both efficiency and accuracy for large-scale applications. This system is designed to effectively identify and resolve coreference links in text, minimizing computational overhead without sacrificing performance. By leveraging advanced neural network architectures, our approach integrates various layers of contextual embeddings and attention mechanisms to ensure high-quality predictions on coreference pairs. Additionally, we implement optimization strategies to enhance the processing speed, facilitating deployment in real-world scenarios. Comprehensive evaluations on benchmark datasets reveal that our model not only improves accuracy metrics compared to existing methods but also maintains a rapid inference time suitable for large-scale text processing. Through rigorous testing, we establish that our system can operate efficiently while delivering precise coreference resolutions, setting a new standard for future developments in this area. Our Contributions. The contributions of this work are as follows: \bullet We propose a novel End-to-End Neural Coreference Resolution system that achieves a harmonious balance between efficiency and accuracy, tailored specifically for large-scale applications. \bullet Our approach utilizes cutting-edge neural network architectures, incorporating contextual embeddings and attention mechanisms for superior prediction quality in identifying coreference links. \bullet Extensive evaluations demonstrate that our model significantly surpasses existing methods in accuracy while maintaining rapid inference times, making it suitable for real-world text processing challenges. 2 Related Work 2.1 Neural Coreference Resolution The incorporation of various techniques and frameworks enhances the performance of coreference resolution systems significantly. A hybrid cache design allows global and local entities to be captured separately, leading to reduced cache misses and improved F1 scores in long document contexts Guo et al. (2023) . Meanwhile, employing centering theory and its transitions in a graphical format aids in refining neural models, despite contextualized embeddings already embedding coherence information Chai and Strube (2022) ; Jiang et al. (2022) . Additionally, reinforcement learning approaches, including actor-critic methods, effectively combine rule-based strategies with neural networks to improve mention clustering and detection Wang and Jin (2022a) ; Wang and Jin (2022b) . Implementing sentence-incremental systems facilitates real-time processing of coreference clusters, outperforming traditional methods Grenander et al. (2023) . In multilingual environments, leveraging synthetic parallel datasets contributes to a consistent performance increase by providing supplementary coreference knowledge Tang and Hardmeier (2023) ; Pražák and Konopík (2024) . Finally, a novel approach focusing on mention annotations alone accelerates domain adaptation processes for coreference models Gandhi et al. (2022) . 2.2 Efficient Large-Scale Systems Recent advancements in large-scale systems have focused on enhancing computational efficiency and performance across various applications. For instance, novel methodologies such as ZeroQuant facilitate post-training quantization of large-scale transformer models, achieving significant speedups without compromising accuracy Yao et al. (2022) . In the realm of gradient optimization, memory-efficient gradient unrolling methods have shown superior performance in bi-level optimization tasks, thereby enhancing scalability Shen et al. (2024) . Additionally, methods like Layerwise Importance Sampled AdamW (LISA) optimize fine-tuning of large language models by applying importance sampling techniques to balance efficiency and performance Pan et al. (2024) . Efforts to streamline robotic 3D reconstruction for visual seafloor mapping further illustrate the emphasis on computationally efficient systems She et al. (2023) . These diverse developments signify a collective aim to optimize performance and efficiency in large-scale computing environments, including platforms for model training and deployment Dolev et al. (2023) ; Fang et al. (2024) . Finally, the introduction of frameworks that utilize retrieval augmentation emphasizes ongoing efforts to refine problem-solving abilities within large models Liu et al. (2024b) . 2.3 Accuracy in NLP Tasks Enhancements in large language models can significantly influence performance, as demonstrated by an improved LoRA fine-tuning algorithm that boosts accuracy, F1 score, and MCC in various NLP tasks Hu et al. (2024) . Moreover, the importance of explainability in model predictions is underscored by methods like COCKATIEL, which provides meaningful insights into neural networks by revealing the concepts utilized for predictions Jourdan et al. (2023) . Memory-augmented transformations further contribute to accuracy in knowledge-intensive tasks by efficiently integrating external knowledge Wu et al. (2022) . In addressing specific linguistic challenges, a morpheme-aware tokenization approach illustrates the potential of linguistically informed strategies to enhance performance, particularly in languages like Korean Jeon et al. (2023)